Faster NTL

[ewx]

NTL appear to have upgraded our internet connection at some point. We now get around 1800Kbit/s download, which is presumably a nominal 2Mbit/s service; uploads top out at around 170Kbit/s. I don't know when this happened, as I don't regularly monitor the speed of our connection.

I still have the measures described here in place, though with the rate limit now at 199Kbit.

Comments

[3c66b.livejournal.com]

Yes, they did the same to Aldabra's NTL a while ago. It's slightly annoying that one has to tweak the tc stuff by hand whenever the speed of the connection changes...

[ewx.livejournal.com]

I couldn't find anything saying what the nominal upload speed was. Sometime I'll try uploading something large without the shaping and see if they still have the original underlying bug.

[pm215]

You could have some kind of funky cron job that tried to automatically determine the best limiting rate...

[ewx.livejournal.com]

..and indeed monitor the download speed.

[sparrowsion]

They're rolling out an increase in speed across the board—I got notification, but then that involved a significant change to the permitted usage. Apparently I can expect a bried period at 1M without the 3G/month cap before they restore me to the current state.

[mikewd]

You can find out the upload speed from the Cable Modem itself using SNMP as it's implemented as bandwidth limiting in the CM.

You can either wade through the DOCSIS MIBs snd do it yourself or parse the output of Robin Walkers DocsDiag program (where it's listed as "QoS max upstream bandwidth").

(I wouldn't say that the CM implementing the bandwidth limit as a simple tail drop FIFO queue is really a "bug" - it's just not useful if you care more about latency than maximising throughput. Whatever mechanism is in use by the ISP you're going to need to implement your own traffic shaping if you want to optimise performance for your requirements.)

[ewx.livejournal.com]

The actual algorithm used, whatever it may be, does not maximize throughput. Proof: the traffic shaping limits did not just improve latency but also throughput.

[(Anonymous)]

RJK is right: if you want to maximise throughput with TCP then you need to minimize the queueing delay so that the TCP endpoints have an accurate model of the connection's uncongested bandwidth and RTT.